diff --git a/tests/v1/core/test_kv_cache_utils.py b/tests/v1/core/test_kv_cache_utils.py index 3c82180474..07c2222145 100644 --- a/tests/v1/core/test_kv_cache_utils.py +++ b/tests/v1/core/test_kv_cache_utils.py @@ -1068,6 +1068,22 @@ def _stats(requests: int, queries: int, hits: int) -> PrefixCacheStats: return PrefixCacheStats(requests=requests, queries=queries, hits=hits) +def test_metrics_empty_distinguishes_no_queries_from_no_hits(): + """`hit_rate` alone cannot tell the two apart; `empty` can. + + Both an unobserved window and a genuine all-miss window report a hit + rate of 0.0, so anything surfacing that number to a human has to check + `empty` first - which is what the prefix-cache log line does. + """ + metrics = CachingMetrics(max_recent_requests=5) + assert metrics.empty + assert metrics.hit_rate == 0.0 + + metrics.observe(_stats(1, 20, 0)) + assert not metrics.empty + assert metrics.hit_rate == 0.0 + + def test_metrics(): """Test the prefix caching metrics.""" metrics = CachingMetrics(max_recent_requests=5) diff --git a/vllm/v1/metrics/loggers.py b/vllm/v1/metrics/loggers.py index fee6679fe7..26a4142df9 100644 --- a/vllm/v1/metrics/loggers.py +++ b/vllm/v1/metrics/loggers.py @@ -296,19 +296,16 @@ class LoggingStatLogger(StatLoggerBase): log_parts.append("Preemptions: %d") log_args.append(self.num_preemptions) - log_parts.extend( - [ - "%s KV cache usage: %.1f%%", - "Prefix cache hit rate: %.1f%%", - ] - ) + log_parts.append("%s KV cache usage: %.1f%%") log_args.extend( [ self.kv_cache_device, self.last_scheduler_stats.kv_cache_usage * 100, - self.prefix_caching_metrics.hit_rate * 100, ] ) + if not self.prefix_caching_metrics.empty: + log_parts.append("Prefix cache hit rate: %.1f%%") + log_args.append(self.prefix_caching_metrics.hit_rate * 100) if envs.VLLM_COMPUTE_NANS_IN_LOGITS: log_parts.append("Corrupted: %d reqs")